Joint disentangled representation and domain adversarial training for EEG-based cross-session biometric recognition in single-task protocols.

Abstract

The increasing adoption of wearable technologies highlights the potential of electroencephalogram (EEG) signals for biometric recognition. However, the intrinsic variability in cross-session EEG data presents substantial challenges in maintaining model stability and reliability. Moreover, the diversity within single-task protocols complicates achieving consistent and generalized model performance. To address these issues, we propose the Joint Disentangled Representation with Domain Adversarial Training (JDR-DAT) framework for EEG-based cross-session biometric recognition within single-task protocols. The JDR-DAT framework disentangles identity-specific features through mutual information estimation and incorporates domain adversarial training to enhance longitudinal robustness. Extensive experiments on longitudinal EEG data from two publicly available single-task protocol datasets-RSVP-based (Rapid Serial Visual Presentation) and MI-based (Motor Imagery)-demonstrate the efficacy of the JDR-DAT framework, with the proposed method achieving average accuracies of 85.83% and 96.72%, respectively.